Loss re-weighting for imbalanced vision, circa TensorFlow 1.14

What it does This repo holds the official TensorFlow code for the CVPR 2019 paper Class-Balanced Loss Based on Effective Number of Samples. It implements re-weighted softmax and focal loss using per-class weights derived from the paper’s “effective number of samples” concept, and includes training and evaluation pipelines for long-tailed CIFAR, iNaturalist, and ImageNet.

The interesting bit The repo goes beyond a minimal loss-function snippet: it bundles full TPU training pipelines forked from Google’s official ResNet-TPU tutorial, plus pre-trained ResNet-50 checkpoints for ImageNet and iNaturalist 2018. That makes it a complete research baseline rather than a toy example.

Key highlights

• Reference TensorFlow implementation of the CVPR 2019 class-balanced loss.
• Training scripts and tfrecords handling for long-tailed CIFAR variants.
• TPU-ready pipelines for iNaturalist 2017/2018 and ImageNet, forked from the official TensorFlow TPU ResNet example.
• Pre-trained ResNet-50 checkpoints provided for ImageNet and iNaturalist 2018.
• Exposes focal loss, class-balanced weighting, and last-layer initialization in the codebase.

Caveats

• Dependencies are pinned to TensorFlow 1.14 and Python 3.6, so the code is showing its age.
• The TPU workflow assumes Google Cloud Storage and Cloud TPU; it is not a self-contained CPU/GPU notebook.
• The README treats the theory lightly—you’ll need the paper to understand what “effective number” actually means.

Verdict Worth cloning if you are reproducing the 2019 baseline or need a long-tailed recognition benchmark to beat. Look elsewhere if you want a modern, drop-in loss layer for PyTorch or newer TensorFlow.